Implementation:SqueezeAILab ETS YAML Config Loading

Knowledge Sources	ETS PyYAML
Domains	Configuration, Experiment_Design
Last Updated	2026-02-14 02:00 GMT

Overview

Concrete tool for loading YAML hyperparameter configuration files using PyYAML's safe_load, as used by the ETS experiment pipeline.

Description

The main() function in rebase.py loads a YAML configuration file specified via --parameter_path using yaml.safe_load(). The resulting dictionary is passed directly to Tree.__init__(), Tree.expand(), Tree.select_and_expand(), and reward_guided_search() as the paras parameter.

Usage

Used at the start of every ETS experiment run. The YAML file must contain all required keys: temperature, max_step_tokens, max_tokens, select_method, num_threads, softmax_temperature, reward_model_type, policy_model_type, store_path, width, lambdac, lambdas.

Code Reference

Source Location

• Repository: ETS
• File: rebase.py
• Lines: 722-723

Signature

import yaml

with open(args.parameter_path, 'r', encoding='utf-8') as file:
    paras = yaml.safe_load(file)

Import

import yaml

I/O Contract

Inputs

Name	Type	Required	Description
parameter_path	str	Yes	Path to YAML config file (passed via --parameter_path CLI arg)

Outputs

Name	Type	Description
paras	dict	Dictionary of hyperparameters consumed by Tree and reward_guided_search

Expected keys in output dict:

Key	Type	Default	Description
temperature	float	1.0	Sampling temperature for text generation
max_step_tokens	int	256	Maximum tokens per reasoning step
max_tokens	int	1024	Maximum total tokens per trajectory
select_method	str	"softmax_costmodel"	Node selection strategy
num_threads	int	8	Number of SGLang parallel threads
softmax_temperature	float	0.2	Temperature for softmax width allocation
reward_model_type	str	"llemma"	Reward model architecture type
policy_model_type	str	"llemma"	Policy model architecture type
store_path	str	—	Directory for per-question result files
width	int	16	Total search budget (number of trajectories)
lambdac	float	1.0	Cost penalty weight in ILP objective
lambdas	float	1.5	Diversity penalty weight (0 disables)

Usage Examples

Example YAML Config (width=16)

temperature: 1.0
max_step_tokens: 256
max_tokens: 1024
select_method: "softmax_costmodel"
num_threads: 8
softmax_temperature: 0.2
reward_model_type: "llemma"
policy_model_type: "llemma"
store_path: ./exp_results/ets_16_math500/
width: 16
lambdac: 1.0
lambdas: 1.5

Loading in Code

import yaml

# Load config from file
with open("hype-parameters/ets_16_math500.yaml", 'r', encoding='utf-8') as file:
    paras = yaml.safe_load(file)

# Access parameters
print(paras["width"])           # 16
print(paras["select_method"])   # "softmax_costmodel"
print(paras["lambdas"])         # 1.5

Related Pages

Implements Principle

• Principle:SqueezeAILab_ETS_Hyperparameter_Configuration

Uses Heuristic

• Heuristic:SqueezeAILab_ETS_Softmax_Temperature_Tuning